LangChain: build a helper for repeated website work

Choose one repeated website job

If you keep opening a work website to check the same information, write down the steps and the result you need. Reading a request's status and preparing a draft summary is a useful starting point. Keep the first exercise simple enough to check by hand.

LangChain's October 29 update to Managed Deep Agents gives custom AI helpers a browser and the ability to delegate longer jobs while continuing a conversation. Managed Deep Agents is a service for building and running your own AI helper through LangSmith. Version 0.10 is available in public beta, an early release for people to try.

Read the announcement: LangChain: October 29 browser and background-helper release.

Work with the page people already use

A website does not always offer a direct software connection for a particular task. The new browser tools let a helper read the page, click, fill a form and inspect what happened. LangChain also describes using them to check an app after a coding change.

The browser runs remotely through LangSmith. It has its own signed-in state and access, so your builder needs to decide which accounts and information it may use. When a suitable direct connection exists, LangChain recommends using that route.

For the status-checking example, ask the builder to show where each result came from. A clear summary should let you compare the answer with the original page and spot information that was missed.

Keep signing in as a human step

When a website needs you to sign in, the helper can hand you control of its browser. You enter your credentials directly on the website, complete any extra verification and hand control back. You do not need to paste your password into the conversation.

While you control that window, the helper's browser tools cannot act on it. Handoff links expire after 15 minutes and should be kept private: someone with a valid link can control the browser.

Your builder can keep a browser within one conversation or reuse a particular person's sign-in across their conversations. Choose that deliberately. Shared sign-in state means switching accounts or signing out can affect that person's other conversations. The helper can still read information the signed-in account is allowed to see.

Keep longer work visible

The update also lets a helper delegate a longer task to another helper working in the background. The main conversation stays available while that work continues, and the result returns when it finishes. The release adds a Stop button for runs in Slack threads.

That can make a tool easier to use, but agreeing what counts as finished still matters. Ask how a person can cancel work, see what happened and check whether a website action succeeded.

Try a limited demonstration first

This building service needs technical setup and a LangSmith account. It also needs a private access key to connect to the AI service you choose. LangSmith lists usage charges, and that AI service can add its own bill. Ask your account owner to confirm access and the full running cost before starting. Beta features and setup requirements can change.

1. Choose one website task and a test account with only the sample information it needs.
2. Write the expected result and the point where a person should take control.
3. Try the normal path, a failed sign-in and a page that is missing information.
4. Check the original page, the draft result and the cost. Confirm you can stop the work.

Bring your knowledge of the work

Our perspective is that describing the repeated task is a valuable part of building the helper. You know what a useful result looks like and where mistakes would matter. Bring that knowledge into a guided build or training session, then use the demonstration to decide what deserves another step.